Machine learning for extracting spatio-temporal biological patterns on evolving domains

我们即将进入一个新时代，在这个时代，荧光活细胞显微镜的进步使人们有可能在几秒钟的时间尺度上获得细胞的详细3D扫描，揭示出以前无法成像的快速生物过程。然而，新的3D数据及其数据量的复杂性给它们的定量分析带来了巨大的挑战。为了应对这些挑战，我们将开发新的机器学习工具，这些工具可以自动检测3D细胞图像中的生物结构，并量化它们如何随时间变化。我们将专注于与不断发展的曲面相关的更易处理和更普遍的问题。不断演变的表面是生物系统跨尺度的一个特征，从迁移或分裂细胞时细胞膜的变形，到胚胎在发育过程中形成的表面。在数学上，这样的表面可以表示为图形，即连接的节点的网格，其中表面上或靠近表面的测量可以是每个节点的特征，例如已经用荧光标记标记的特定分子的丰度。引入图的概念具有变革性，因为最近高效的图神经网络(GNN)推动了图上特征的自动分类。我们将在最近的工作基础上，开发出一种用于在3D图像中检测生物表面的高精度概念验证方法。为了在高吞吐量环境中使用，该方法需要在计算上变得更高效，这将通过使用快速发展的、功能强大的TensorFlow框架来实现高度并行的GPU编程。为了在真实的实验数据上验证这些方法，我们将研究与细胞饮用相关的细胞表面变形，以及斑马鱼肌肉片段形成过程中更复杂的细胞伸长和融合的多细胞问题。下一个主要步骤将是生成随时间变化的生物表面的图形表示。这里的挑战是在随后的时间点匹配图表上的相应节点，请注意，尽管时间分辨率很高，但一个时间点和下一个时间点之间的变形可能很大。初步结果表明，用于图形匹配的谱方法通常难以处理较大的表面变形，可以通过结合诸如曲面曲率之类的附加特征来改进。在这里，我们还建议结合荧光蛋白在细胞膜内或靠近细胞膜的空间分布，以提高表面匹配的准确性。除了前面的生物表面的例子，我们还将包括接近球形的早期斑马鱼胚胎表面的例子，并研究细胞形状变化在神经系统前体前神经板形成中的作用。为了分析表面上依赖时间的特征，我们将引入新的方法来随时间演变图形结构，保持节点数量恒定，这是执行分类和特征提取任务时所需的。特征的监督学习需要人工标注，而对于动态特征则很难获得。我们将开发新的图形界面，并使用虚拟现实技术，让研究人员以直观的方式与数据交互并对数据进行注释。我们将要开发的自动特征检测工具将帮助生物学家获得更详细的高质量数据，这些数据可以用来通过比较健康和疾病的细胞来研究生物机制，或者研究药物的效果。这些工具的开发将与实验者和更广泛的生物成像用户社区密切合作，与伦敦的克里克研究所和MRC LMS三家主要成像机构以及利物浦大学合作进行。

We are arriving in a new era where advances in fluorescence live cell microscopy make it possible to acquire detailed 3D scans of cells on a timescale of seconds, revealing fast biological processes that could not be imaged before. The complexity of the new 3D data and their volume, however, pose significant challenges for their quantitative analysis. To address these challenges, we will develop novel machine learning tools that can automatically detect biological structures within the 3D cell images and quantify how they change over time. We will focus on more tractable and generalisable problems that are associated with evolving surfaces. Evolving surfaces are a feature of biological systems across scales, from deformations of the cell membrane in migrating or dividing cells, to the surface of an embryo that takes on shape during development. Mathematically, such surfaces can be represented as a graph, a mesh of connected nodes, where measurements on or close to the surface can be a feature of each node, for example the abundance of a particular molecule that has been tagged with a fluorescent marker. Introducing the concept of graphs is transformative, because automated classification of features on graphs has recently been propelled by highly efficient graph neural networks (GNNs). We will build on recent work where we developed a highly accurate proof-of-concept method for detecting biological surfaces in 3D images. To be used in a high-throughput environment this method needs to become computationally more efficient, which will be achieved by employing the rapidly evolving, powerful Tensorflow framework for highly parallel GPU-programming. To validate these methods on real experimental data, we will investigate cell surfaces deformation associated with cell drinking, and the more complex multicellular problem of cell elongation and fusion during zebrafish muscle segment formation. The next major step will be to generate graph representations of biological surfaces that change over time. The challenge here is to match corresponding nodes on the graph at subsequent timepoints, noting that despite a high temporal resolution, deformations between one timepoint and the next can be large. Preliminary results show that spectral methods for graph matching, which normally struggle with large surface deformations, can be improved by incorporating additional features such as curvature of the surface. Here we propose to also incorporate spatial distributions of fluorescent proteins in or close to the cell membrane to improve the accuracy of surface matching. In addition to the previous examples of biological surfaces, we will include that of the early zebrafish embryo surface which is near spherical, and investigate the role of cell shape changes in the formation of the anterior neural plate, a precursor of the nervous system. For analysing time dependent features on surfaces, we will be introducing novel methods to evolve the graph structure over time, keeping the number of nodes constant, as is required when performing classification and feature extraction tasks. Supervised learning of features requires manual annotations, which for dynamic features are difficult to obtain. We will develop new graphical interfaces and use virtual reality technology to allow researchers to interact with and annotate the data in an intuitive manner. The tools for automated feature detection that we are going to develop will help biologists to obtain more detailed high-quality data, which can be used to investigate biological mechanisms by comparing healthy with diseased cells, or to study the effects of drugs. The development of these tools will be undertaken in close collaboration with experimentalists and the wider community of bioimaging users, in partnership with three major imaging facilities, the Crick Institute and MRC LMS in London, and at Liverpool University.